Skip to content

[NoQA] Replace err-on-side-of-flagging with self-critique gate in code reviewer#85068

Closed
kacper-mikolajczak wants to merge 2 commits intoExpensify:mainfrom
kacper-mikolajczak:suppress-self-contradicting-reviews
Closed

[NoQA] Replace err-on-side-of-flagging with self-critique gate in code reviewer#85068
kacper-mikolajczak wants to merge 2 commits intoExpensify:mainfrom
kacper-mikolajczak:suppress-self-contradicting-reviews

Conversation

@kacper-mikolajczak
Copy link
Contributor

@kacper-mikolajczak kacper-mikolajczak commented Mar 12, 2026

Explanation of Change

The AI code reviewer currently has an EXCEPTION clause telling it to "err on the side of including" uncertain violations. This causes ~17-23% of review comments to contain visible self-contradiction - the model flags a violation then talks itself out of it in the same comment ("actually this is correct", "upon re-reading, this is fine").

A previous attempt (PR #83184) added a "reality check" instruction that told the model to re-read and confirm violations before posting. This was reported to have the opposite effect - making the model deliberate more visibly in its output.

This PR replaces the EXCEPTION clause with a "self-critique gate" instruction that:

  • Tells the model to silently verify each violation before including it
  • Omit any violation where there is doubt (rather than include with caveats)
  • Never include hedging language ("actually", "upon re-reading", "wait") in violation bodies
  • Frame uncertain cases as "omit" rather than "check if correct" - avoiding the deliberation trigger

The key behavioral shift: "omit if not confident" instead of "check if correct."

Companion PRs: Auth and Web-Expensify (same change adapted for inline comment posting).

Fixed Issues

$ https://github.com/Expensify/Expensify/issues/605351
PROPOSAL:

Tests

  1. Trigger the AI reviewer on a PR with code that is borderline (could be flagged but is actually correct)
  2. Verify the reviewer does not post self-contradicting comments
  3. Verify that genuine violations are still flagged with clear, definitive language
  4. Verify no hedging phrases ("actually", "upon re-reading", "wait", "however") appear in posted comments

Offline tests

N/A - AI reviewer agent prompt, no offline behavior.

QA Steps

// TODO: These must be filled out, or the issue title must include "[No QA]."

N/A - AI reviewer agent configuration change. No user-facing UI changes.

  • Verify that no errors appear in the JS console

PR Author Checklist

  • I linked the correct issue in the ### Fixed Issues section above
  • I wrote clear testing steps that cover the changes made in this PR
    • I added steps for local testing in the Tests section
    • I added steps for the expected offline behavior in the Offline steps section
    • I added steps for Staging and/or Production testing in the QA steps section
    • I added steps to cover failure scenarios (i.e. verify an input displays the correct error message if the entered data is not correct)
    • I turned off my network connection and tested it while offline to ensure it matches the expected behavior (i.e. verify the default avatar icon is displayed if app is offline)
    • I tested this PR with a High Traffic account against the staging or production API to ensure there are no regressions (e.g. long loading states that impact usability).
  • I included screenshots or videos for tests on all platforms
  • I ran the tests on all platforms & verified they passed on:
    • Android: Native
    • Android: mWeb Chrome
    • iOS: Native
    • iOS: mWeb Safari
    • MacOS: Chrome / Safari
  • I verified there are no console errors (if there's a console error not related to the PR, report it or open an issue for it to be fixed)
  • I followed proper code patterns (see Reviewing the code)
    • I verified that any callback methods that were added or modified are named for what the method does and never what callback they handle (i.e. toggleReport and not onIconClick)
    • I verified that comments were added to code that is not self explanatory
    • I verified that any new or modified comments were clear, correct English, and explained "why" the code was doing something instead of only explaining "what" the code was doing.
    • I verified any copy / text shown in the product is localized by adding it to src/languages/* files and using the translation method
      • If any non-english text was added/modified, I used JaimeGPT to get English > Spanish translation. I then posted it in #expensify-open-source and it was approved by an internal Expensify engineer. Link to Slack message:
    • I verified all numbers, amounts, dates and phone numbers shown in the product are using the localization methods
    • I verified any copy / text that was added to the app is grammatically correct in English. It adheres to proper capitalization guidelines (note: only the first word of header/labels should be capitalized), and is either coming verbatim from figma or has been approved by marketing (in order to get marketing approval, ask the Bug Zero team member to add the Waiting for copy label to the issue)
    • I verified proper file naming conventions were followed for any new files or renamed files. All non-platform specific files are named after what they export and are not named "index.js". All platform-specific files are named for the platform the code supports as outlined in the README.
    • I verified the JSDocs style guidelines (in STYLE.md) were followed
  • If a new code pattern is added I verified it was agreed to be used by multiple Expensify engineers
  • I followed the guidelines as stated in the Review Guidelines
  • I tested other components that can be impacted by my changes (i.e. if the PR modifies a shared library or component like Avatar, I verified the components using Avatar are working as expected)
  • I verified all code is DRY (the PR doesn't include any logic written more than once, with the exception of tests)
  • I verified any variables that can be defined as constants (ie. in CONST.ts or at the top of the file that uses the constant) are defined as such
  • I verified that if a function's arguments changed that all usages have also been updated correctly
  • If any new file was added I verified that:
    • The file has a description of what it does and/or why is needed at the top of the file if the code is not self explanatory
  • If a new CSS style is added I verified that:
    • A similar style doesn't already exist
    • The style can't be created with an existing StyleUtils function (i.e. StyleUtils.getBackgroundAndBorderStyle(theme.componentBG))
  • If new assets were added or existing ones were modified, I verified that:
    • The assets are optimized and compressed (for SVG files, run npm run compress-svg)
    • The assets load correctly across all supported platforms.
  • If the PR modifies code that runs when editing or sending messages, I tested and verified there is no unexpected behavior for all supported markdown - URLs, single line code, code blocks, quotes, headings, bold, strikethrough, and italic.
  • If the PR modifies a generic component, I tested and verified that those changes do not break usages of that component in the rest of the App (i.e. if a shared library or component like Avatar is modified, I verified that Avatar is working as expected in all cases)
  • If the PR modifies a component related to any of the existing Storybook stories, I tested and verified all stories for that component are still working as expected.
  • If the PR modifies a component or page that can be accessed by a direct deeplink, I verified that the code functions as expected when the deeplink is used - from a logged in and logged out account.
  • If the PR modifies the UI (e.g. new buttons, new UI components, changing the padding/spacing/sizing, moving components, etc) or modifies the form input styles:
    • I verified that all the inputs inside a form are aligned with each other.
    • I added Design label and/or tagged @Expensify/design so the design team can review the changes.
  • If a new page is added, I verified it's using the ScrollView component to make it scrollable when more elements are added to the page.
  • I added unit tests for any new feature or bug fix in this PR to help automatically prevent regressions in this user flow.
  • If the main branch was merged into this PR after a review, I tested again and verified the outcome was still expected according to the Test steps.

Screenshots/Videos

Android: Native
Android: mWeb Chrome
iOS: Native
iOS: mWeb Safari
MacOS: Chrome / Safari

@kacper-mikolajczak
Replace err-on-side-of-flagging with self-critique gate in code reviewer
76cb78e 
The EXCEPTION clause telling the model to include uncertain violations
was contributing to self-contradicting review comments (17-23% of output).
Replace it with a self-critique gate that instructs the model to silently
omit any violation it is not confident about, and to never include
deliberation or hedging language in comment bodies.
@kacper-mikolajczak
Copy link
Contributor Author

kacper-mikolajczak commented Mar 18, 2026

Closing - this approach re-invents prior attempts and doesn't target the root cause. Will investigate a different approach.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@kacper-mikolajczak kacper-mikolajczak

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@kacper-mikolajczak